2026-11-13: why the momentum picks lost while the market rose
Question (owner, 2026-11-13): the momentum sleeve's stock picks are down 12.4% since the 2026-06-11 clean clock while SPY is up 4.8%. Why, and can it be fixed?
Answer. The software did what the strategy says. The strategy buys the 50 stocks that rose the most, and the most steadily, over the past year. Twice in this window, in July and in October, the market turned on exactly those stocks while the rest of the market went up. No part of the loss traces to a malfunction, so there is nothing to repair in the code. Changing the picks means changing the strategy, which is an owner decision under the SHIP gate, and nothing here is a case for doing it. No change was made to the sleeve.
All figures are simulated paper trading. Reproduce with
python research/momentum_picks_diagnosis.py --prices <copy of the trading PC's data/raw>
(read-only; prices through 2026-11-12, state as of commit 724638d).
Three checks
1. Did it pick the stocks the strategy says to pick? Yes.
The top 50 was recomputed from prices with separate code and compared with the targets each live selection committed.
| Selection | Live targets that are in the recomputed top 50 |
|---|---|
| 2026-06-11 | 47 of 50 |
| 2026-07-01 | 49 of 50 |
| 2026-08-03, 09-01, 10-01, 11-03 | 50 of 50 |
The 2026-11-02 selection, which ranked only 37 names, was voided and redone on 11-03
(2026-11-03_momentum_selection_voided.md). It is not counted here.
2. Did the account hold what was targeted? Yes, within about a point.
Replaying the committed targets day by day, held between the monthly selections:
| Committed targets | Live stocks-only record | |
|---|---|---|
| 2026-06 (from 06-12) | +0.6% | −0.0% |
| 2026-07 | −9.0% | −9.3% |
| 2026-08 | +1.0% | +0.6% |
| 2026-09 | +0.7% | +0.5% |
| 2026-10 | −1.0% | −1.6% |
| 2026-11 (to 11-12) | −2.7% | −2.1% |
| 107 days | −10.3% | −11.6% |
Daily correlation is 0.81, and 0.90 on two-day sums (the live record is marked at run time, not at
the close). The 1.3-point difference over five months is the cost of not sitting exactly on the
targets: small leftover positions, fill timing and spreads. On 11-13 the broker account held $103
of its $2,052 of stock in 28 names that are no longer targets, which is the leftover class FIN-2 and
PBC-1 already ruled on. One of the 28 is ECHO ($20), the renamed SATS: it has no price file under
its new ticker, so the daily run logs it as untradeable and skips it (the caveat
_liquidate_to_cash documents). The published −12.4% runs one more day (11-13: picks −0.9%, SPY
+0.8%).
The share of the account in stocks also matches the design: 31% on average over the window, against a mean of 31% across the saved backtest's 194 selections (5th to 95th percentile: 9% to 49%).
3. Was it the picks or the market? The signal pointed the wrong way.
Close of each selection day to the close of the next. "Picks" uses the committed weights.
| Held from | Picks | Same picks, equal weight | Average stock | SPY | Signal's best tenth | Signal's worst tenth | Equal-weight picks beat this share of 3,000 random 50-stock baskets |
|---|---|---|---|---|---|---|---|
| 06-11 | −2.8% | +0.9% | +2.0% | +1.3% | −0.9% | +3.9% | 20% |
| 07-01 | −4.7% | −12.0% | +2.3% | +1.6% | −6.6% | +9.4% | 0% |
| 08-03 | −0.4% | +0.1% | −0.8% | +0.5% | −1.6% | +4.4% | 75% |
| 09-01 | +1.9% | +7.6% | −4.0% | +0.5% | +4.4% | −9.0% | 100% |
| 10-01 | −3.9% | −12.5% | +5.8% | −1.3% | −8.0% | +11.3% | 0% |
| 11-03 (9 days) | −0.9% | −2.9% | +0.3% | +0.4% | −1.3% | +0.2% | 0% |
| Whole window | −10.4% | −18.7% | +5.4% | +3.1% | −13.5% | +20.6% |
- Across the whole universe of 884 stocks, the tenth the signal ranked best lost 13.5% and the tenth it ranked worst gained 20.6%. Past winners fell and past losers rose. That is a momentum reversal, and it is the known way this kind of strategy loses.
- The loss is two months. From the July and October selections the equal-weighted picks finished below every one of 3,000 random baskets. From the September selection they finished above every one. The picks are a concentrated bet on one thing, so they land at one end or the other.
- The weighting did its job. HRP (the rule that gives calmer stocks more of the money) turned an 18.7% loss at equal weights into 10.4%.
- The negative-control sleeve, which buys recent losers, shows the same market from the other side: its picks are up 11.5% since 2026-07-13.
How unusual is it?
The same measure, picks minus the average stock, over the saved backtest's 193 monthly periods (2010 to early 2026):
| Backtest, 16 years | This window | |
|---|---|---|
| Worst single month, committed weights | −9.1% (2023-01) | −9.7% (October), −7.0% (July) |
| 1st percentile month | −6.5% | |
| Worst single month, equal weights | −7.5% (2021-03) | −18.3% (October), −14.3% (July) |
| Worst five-month stretch | −12.8% | −15.8% |
October was worse than any month in the backtest, and the five months together were worse than any five-month stretch in it. That fits two readings and this data cannot separate them: a rare market event, or a backtest that understated how badly the strategy can lose. The second is the concern the overfitting tests already carry (PBO 62.5%; the headline Sharpe is the best of eight start-day schedules).
What this does not show
- Five monthly periods are a small sample. The decay judge reads
OKonly because it cannot tell five bad months from noise (FIN-3). - Both eras use today's index membership, so both overstate what the average stock earned. The comparison is like for like, not exact.
- The replay assumes each selection is filled at that day's close. Live fills came hours earlier or later, which is why single days differ while months agree.
- The price copy was taken from the trading PC on 2026-11-13. A later re-download re-bases adjusted prices and can move these figures slightly.
Decision
- No change to the sleeve, its config or its state. Strategy, construction and sizing values are owner-only and go through the SHIP gate. Retuning a rule against the five months it just lost is fitting to noise.
- The December gate already covers the real-money question.
thales go-live-gatereads NOT PASSED today: forward Sharpe −1.42, 95% lower bound −4.47, where the pre-registered rule needs the lower bound above 0. It has 108 of the 126 clean days it needs and reaches 126 around 2026-12-10. - Nothing new is filed. The leftover positions are FIN-2 and PBC-1, already ruled.
Files
research/momentum_picks_diagnosis.py: the script behind every figure above.web/public/data/momentum.json(investedblock): the live stocks-only record it is compared with.results/weights_history.parquet,data/raw(frozen 2026-06-09): the backtest-era comparison.